Superpixel-enhanced Deep Neural Forest for Remote Sensing Image Semantic Segmentation

NOTE: The CODE is UNDER maintenance since 13 Oct 2020. Codes and modifications will continue to be updated.

Results for Paper: Superpixel-enhanced Deep Neural Forest for Remote Sensing Image Semantic Segmentation

Environments

• Python 3.6.2
• Tensorflow 1.6.0
• Numpy 1.13.1
• Opencv-python
• Matplotlib
• Scipy

Data

• Download the test image (RGB for Potsdam/IRRG for Vaihingen) and RGB label image (Fully Reference/No Boundary) from ISPRS 2D semantic labelling website.
• Transfer the RGB label image to the corresponding label image (provided).

	Index	R	G	B
Imp	0	255	255	255
Build	1	0	0	255
Low	2	0	255	255
Tree	3	0	255	0
Car	4	255	255	0
Cluster	5	255	0	0
Un	6	0	0	0

• Rename the testing image and label image.

Pre-trained Model

• Download the Pre-trained resnet_v2_101.ckpt and put it into /Code/net/resnet.
• Download the Pre-trained model for Vaihingen Dataset and put it into /Models/Vaihingen.
• Download the Pre-trained model for Potsdam Dataset and put it into /Models/Potsdam.

Evaluation

import predict_potsdam
predict_potsdam.process()

import predict_vaihingen
predict_vaihingen.process()

• The results include the predict RGB image, the predict Label image and the results.txt for accuracy.
• The whole evaluation process is about 20min.

Results

	Imp.S.	Imp.S.	Build.	Build.	Low.V.	Low.V.	Tree	Tree	Car	Car	Mean	Mean	OA
	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU
Potsdam	93.5	87.7	96.3	93.0	89.8	81.5	92.7	86.4	96.7	93.6	93.8	88.4	92.1
Vaihingen	93.6	87.9	96.2	92.6	88.0	78.6	92.6	86.3	85.3	74.4	91.1	83.9	92.6

Acknowledgements

Our code is developed based on：

ssn_superpixels
pytorch_ssn
fully-differentiable-deep-ndf-tf
Neural-Decision-Forests
tensorflow-deeplab-v3
deeplabv3-Tensorflow

Cite

@article{Li2020Superpixel,
title={Superpixel-enhanced deep neural forest for remote sensing image semantic segmentation},
author={Li Mi and Zhenzhong Chen},
journal={ISPRS Journal of Photogrammetry and Remote Sensing},
volume={159},
pages={140-152},
year={2020},
}